我正在尝试识别足球场的中场椭圆。我尝试了在网上找到的不同解决方案，如霍夫变换和模板匹配，但不幸的是，我无法使它们工作。图像取自真实的足球比赛，所以有不同的相机镜头和方向。
目前，我获得了最好的结果，过滤掉了音高之外的像素（使用颜色遮罩），并使用了一些转换，如高斯模糊，Canny，以及边缘轮廓，以及来自openCV的fitEllipse（）等技术。
蒙版后的起始图像：

Canny图像：

在这里，我执行了边缘轮廓，并使用aspect_ratio过滤掉不相关的边缘：

contours, _ = cv2.findContours(canny_image, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

min_aspect_ratio = 3  # Minimum aspect ratio threshold (adjust as needed)
max_aspect_ratio = 50  # Maximum aspect ratio threshold (adjust as needed)

# Create a copy of the original image to draw bounding rectangles
image_with_rectangles = image.copy()

# Iterate through the contours
for contour in contours:
    # Calculate the aspect ratio of the bounding rectangle
    x, y, w, h = cv2.boundingRect(contour)
    aspect_ratio = float(w) / h

    # Check if the aspect ratio falls within the specified range
    if min_aspect_ratio < aspect_ratio < max_aspect_ratio:
        # Draw a green bounding rectangle around the contour
        cv2.rectangle(image_with_rectangles, (x, y), (x + w, y + h), (255, 255, 0), 2)

# Display the image with the bounding rectangles
print("Image with Bounding Rectangles")
cv2_imshow(image_with_rectangles)

filtered_contours = [
    cnt for cnt in contours
    if min_aspect_ratio < (float(cv2.boundingRect(cnt)[2]) / cv2.boundingRect(cnt)[3]) < max_aspect_ratio
]

result_image = np.zeros_like(original_image)  # Create a blank canvas

cv2.drawContours(result_image, filtered_contours, -1, (0, 255, 0), 2)  # Draw the filtered contours in green

# Display or save the result_image as needed
cv2_imshow(result_image)

带边框的图像：

图像轮廓：

最后一步包括尝试拟合椭圆，我遇到了很多问题，使cv2.fitEllipse（）工作，这里我的解决方案：

gray_image = cv2.cvtColor(result_image, cv2.COLOR_BGR2GRAY)
cv2_imshow(gray_image)

# find the contours
contours,hierarchy = cv2.findContours(gray_image, cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)
print("\nNumber of contours detected:", len(contours))

ellipse_image = original_image.copy()
ellipses = []
orientations = []

for c in filtered_contours:

  #if it contains at least 5 points (higher the value lower the number of ellipses)
  if len(c) >= 10:
    print(len(c))

    ellipse = cv2.fitEllipse(c)

    # Get the major and minor axes of the ellipse
    major_axis, minor_axis = ellipse[1]
    # Calculate the aspect ratio of the ellipse
    aspect_ratio = major_axis / minor_axis
    #we want a high aspect ratio, meaning that is much wider then taller
    if 0.1 <= aspect_ratio <= 0.2:
      print("\nRatio: ", aspect_ratio)
      print("M and m axes: ", major_axis, minor_axis)

      if major_axis < minor_axis:
        ellipses.append(ellipse)
        orientations.append(int(ellipse[2]))

for e in ellipses:
   cv2.ellipse(ellipse_image, e, (255, 255, 0), 2)

print("\nMiddlefield contour:")
cv2_imshow(ellipse_image)

最终结果：

如何保持只有一个椭圆（正确的一个）？

VITOR发布后编辑

为了在输入图像上检测到更多的轮廓，需要进行大量的图像预处理操作，如果你能得到一个良好的、干净的中场椭圆形状，那么他的算法就可以只检测和保留一个椭圆。结果有一些假阳性，但是对椭圆的面积、长宽比和长度施加了一些范围限制，我得到了很好的结果。
下面是一些例子。正如你所看到的，这取决于过滤后的轮廓。
好结果：

不太坏的结果：

错误结果（未检测到椭圆）：